SELECTIONS OF MY RESEARCH WORK

Deep Learning Detection of Left Ventricular Wall Motion Abnormality from Volume-Rendered 4DCT

Check here for the GitHub Repository and here for the conference talk!
There has been no deep learning (DL) methodology to analyze 3D LV wall motion abnormality (WMA) from 4DCT due to two pain points. First, a 4DCT dataset is too large (Gigabytes) to fit into the current GPU. Second, 3D motion vector field estimation is complicated and computationally expensive.
A Volume Rendering video (check here for an example showing WMA at the septal wall) solves these two pain points. First, the high-resolution 3D wall motion is represented in the video by a very small data size (Kilobytes) suitable for deep learning experiments. Second, the motion irregularity is encoded by video features such as surface motion or color pattern change which can be simply and intuitively learned by a DL model.
We have designed an in-house program to automatically generate 6 volume rendering videos of the LV blood pool with 6 different views around the LV long axis for each 4DCT study. We have also proposed a deep learning framework to detect the WMA from the video. Our approach as the first DL methodology to detect WMA from 4DCT achieved high detection performance both in cross-validation and testing cohorts (Figure 2).

Deep Learning Chamber Segmentation and Cardiac Imaging Plane Prediction

Check here for some videos of our deep learning products and here for the GitHub Repository!
There are two essential image processing required for cardiac function analysis in CT. Accurate cardiac chamber segmentation is required to do quantitative assessments. Standard cardiac imaging planes such as multiple long-axis (LAX) planes and one short-axis (SAX) stack are required to visualize regional wall motion abnormalities. Doing the processing manually is time-consuming, requires extensive experience, requires specialized software, and leads to inter-reader variability.
We have invented a deep learning (DL) framework to accomplish these two processing tasks. We modified the classical U-Net architecture (Figure 2) for multi-task learning.
We showed the high performance of our DL framework both in segmentation (Figure 3) and plane prediction (Figure 4). For plane prediction, the final deep learning product we can provide to physicians is shown in Figure 1 as a matrix of image panels containing LAX and SAX views. This product is an excellent tool to visualize regional wall motion abnormalities.

Quantification of Regional Cardiac Function - CT SQUEEZ

Assessment of regional myocardial function is critical to the evaluation of both ischemic and non-ischemic cardiomyopathy. Modern multi-detector 4DCT can provide an excellent evaluation of cardiac function by acquiring a series of dose-modulated images spanning the cardiac cycle within a single heartbeat. We have developed a technique called CT Stretch Quantifier for Endocardial Engraved Zones (CT SQUEEZ) to quantify the regional myocardial function in high-resolution volumetric cardiac CT images using a fast, nonrigid, surface registration algorithm that matches geometric features of the surface over time.
As shown in Figure 1, the endocardial surface is represented by triangular meshes. We use a registration algorithm to warp a template mesh at the template frame (end-diastole) to a target mesh at any systolic frame. Once the one-to-one correspondence is established, the regional shortening of the endocardium is represented by the area change of the same mesh at two different frames.
We have implemented this technique in several pilot clinical studies as follows. First, we leveraged SQUEEZ to define an strain range for normal LV (Figure 2). Second, we used CT SQUEEZ to characterize the regional RV dysfunction in patients with congenital heart disease (CHD) (Figure 3). Third, I am now leading a study to define the segment-specific SQUEEZ threshold that leads to accurate detection of LV wall motion abnormality (WMA) (Figure 4).

Precise Quantification of Coronary Stenosis Dimension with CCTA

Coronary CT angiography (CCTA) has emerged as a promising noninvasive method for the visualization of coronary artery disease (CAD). However, It remains challenging to accurately quantify severe stenosis with its dimension usually below the CT image resolution. The conventional Full-Width Half Maximum (FWHM) is only accurate for large vessels but underestimates vessels below the resolution.
We have discovered a relationship between the intra-luminal maximum intensity and the vessel diameter plotted as the vessel intensity calibration curve (Figure 1). We further explained the relation mathematically as IMVV(D)/Io = 1 - k x PSF(D), where IMVV is maximum intensity for vessel diameter D, Io is intensity for large vessels, PSF is a 2D Gaussian Point Spread Function, k is a coefficient related to std of PSF. We have shown the dependency of the calibration curve on the image acquisition settings that change PSF such as focal spot size and recon kernels (Figure 1).
We leveraged this relationship to develop an innovative quantification algorithm based on vessel maximum intensity. We proposed to measure large vessel diameter (> image resolution) by FWHM while measuring small vessel diameter (< image resolution) by vessel intensity calibration. This algorithm was validated to have high accuracy (~95%) to capture the true dimension of stenosis (Figure 2) on the 3D printed anthropomorphic stenosis phantom (Figure 3).